Deep Learning-Based Classification of Ischemic Stroke Using Brain CT Scans

Abstract

The timely clinical intervention of ischemic stroke in brain CT scans requires the early and accurate identification of its presence in the brain but this is not easy as the imaging characteristics are subtle. This paper demonstrates a well-validated deep learning model with architecture-based DenseNet121 and ResNet18 to identify an ischemic stroke on a large-scale CT dataset of 6,653 scans extensively enhanced to approximately 20,000 images to deal with the issue of class imbalance. In addition to using popular CNN models, our work innovates the field with the use of strict k-fold cross-validation and home-based test validation, which guarantees the strength of the results as well as their successful generalization. Both models are highly accurate (DenseNet121: 98.20%, ResNet18: 97.97%) and compete well with the current state-of-the-art methods. The findings indicate relevant clinical applicability, which provides the possibility to provide quick and dependable automated stroke diagnostics in the emergency department, which may benefit clinicians and decrease diagnostic time and enhance patient outcomes. This paper provides a standard of a proven CNN-based stroke classification system and indicates the future research to improve clinical integration.